11 struct io_completion_data {
14 int error; /* output */
15 unsigned long bytes_done[2]; /* output */
16 struct timeval time; /* output */
20 * The ->file_map[] contains a map of blocks we have or have not done io
21 * to yet. Used to make sure we cover the entire range in a fair fashion.
23 static int random_map_free(struct fio_file *f, const unsigned long long block)
25 unsigned int idx = RAND_MAP_IDX(f, block);
26 unsigned int bit = RAND_MAP_BIT(f, block);
28 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
30 return (f->file_map[idx] & (1 << bit)) == 0;
34 * Mark a given offset as used in the map.
36 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
38 unsigned int min_bs = td->o.rw_min_bs;
39 struct fio_file *f = io_u->file;
40 unsigned long long block;
41 unsigned int blocks, nr_blocks;
43 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
44 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
48 unsigned int this_blocks, mask;
49 unsigned int idx, bit;
52 * If we have a mixed random workload, we may
53 * encounter blocks we already did IO to.
55 if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
61 idx = RAND_MAP_IDX(f, block);
62 bit = RAND_MAP_BIT(f, block);
64 fio_assert(td, idx < f->num_maps);
66 this_blocks = nr_blocks;
67 if (this_blocks + bit > BLOCKS_PER_MAP)
68 this_blocks = BLOCKS_PER_MAP - bit;
70 if (this_blocks == BLOCKS_PER_MAP)
73 mask = ((1U << this_blocks) - 1) << bit;
75 f->file_map[idx] |= mask;
76 nr_blocks -= this_blocks;
77 blocks += this_blocks;
81 if ((blocks * min_bs) < io_u->buflen)
82 io_u->buflen = blocks * min_bs;
85 static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
88 unsigned long long max_blocks;
89 unsigned long long max_size;
92 * Hmm, should we make sure that ->io_size <= ->real_file_size?
94 max_size = f->io_size;
95 if (max_size > f->real_file_size)
96 max_size = f->real_file_size;
98 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
106 * Return the next free block in the map.
108 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
109 enum fio_ddir ddir, unsigned long long *b)
111 unsigned long long min_bs = td->o.rw_min_bs;
114 i = f->last_free_lookup;
115 *b = (i * BLOCKS_PER_MAP);
116 while ((*b) * min_bs < f->real_file_size &&
117 (*b) * min_bs < f->io_size) {
118 if (f->file_map[i] != (unsigned int) -1) {
119 *b += ffz(f->file_map[i]);
120 if (*b > last_block(td, f, ddir))
122 f->last_free_lookup = i;
126 *b += BLOCKS_PER_MAP;
130 dprint(FD_IO, "failed finding a free block\n");
134 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
135 enum fio_ddir ddir, unsigned long long *b)
137 unsigned long long r;
141 r = os_random_long(&td->random_state);
142 dprint(FD_RANDOM, "off rand %llu\n", r);
143 *b = (last_block(td, f, ddir) - 1)
144 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
147 * if we are not maintaining a random map, we are done.
149 if (!file_randommap(td, f))
153 * calculate map offset and check if it's free
155 if (random_map_free(f, *b))
158 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
163 * we get here, if we didn't suceed in looking up a block. generate
164 * a random start offset into the filemap, and find the first free
169 f->last_free_lookup = (f->num_maps - 1) *
170 (r / (OS_RAND_MAX + 1.0));
171 if (!get_next_free_block(td, f, ddir, b))
174 r = os_random_long(&td->random_state);
178 * that didn't work either, try exhaustive search from the start
180 f->last_free_lookup = 0;
181 return get_next_free_block(td, f, ddir, b);
185 * For random io, generate a random new block and see if it's used. Repeat
186 * until we find a free one. For sequential io, just return the end of
187 * the last io issued.
189 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
191 struct fio_file *f = io_u->file;
192 unsigned long long b;
193 enum fio_ddir ddir = io_u->ddir;
195 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
196 td->ddir_nr = td->o.ddir_nr;
198 if (get_next_rand_offset(td, f, ddir, &b)) {
199 dprint(FD_IO, "%s: getting rand offset failed\n",
204 if (f->last_pos >= f->real_file_size) {
205 if (!td_random(td) ||
206 get_next_rand_offset(td, f, ddir, &b)) {
207 dprint(FD_IO, "%s: pos %llu > size %llu\n",
208 f->file_name, f->last_pos,
213 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
216 io_u->offset = b * td->o.ba[ddir];
217 if (io_u->offset >= f->io_size) {
218 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
219 io_u->offset, f->io_size);
223 io_u->offset += f->file_offset;
224 if (io_u->offset >= f->real_file_size) {
225 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
226 io_u->offset, f->real_file_size);
233 static inline int is_power_of_2(unsigned int val)
235 return (val != 0 && ((val & (val - 1)) == 0));
238 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
240 const int ddir = io_u->ddir;
241 unsigned int uninitialized_var(buflen);
242 unsigned int minbs, maxbs;
245 minbs = td->o.min_bs[ddir];
246 maxbs = td->o.max_bs[ddir];
251 r = os_random_long(&td->bsrange_state);
252 if (!td->o.bssplit_nr[ddir]) {
253 buflen = 1 + (unsigned int) ((double) maxbs *
254 (r / (OS_RAND_MAX + 1.0)));
261 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
262 struct bssplit *bsp = &td->o.bssplit[ddir][i];
266 if (r <= ((OS_RAND_MAX / 100L) * perc))
270 if (!td->o.bs_unaligned && is_power_of_2(minbs))
271 buflen = (buflen + minbs - 1) & ~(minbs - 1);
274 if (io_u->offset + buflen > io_u->file->real_file_size) {
275 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
283 static void set_rwmix_bytes(struct thread_data *td)
288 * we do time or byte based switch. this is needed because
289 * buffered writes may issue a lot quicker than they complete,
290 * whereas reads do not.
292 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
293 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
296 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
301 r = os_random_long(&td->rwmix_state);
302 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
303 if (v <= td->o.rwmix[DDIR_READ])
310 * Return the data direction for the next io_u. If the job is a
311 * mixed read/write workload, check the rwmix cycle and switch if
314 static enum fio_ddir get_rw_ddir(struct thread_data *td)
318 * Check if it's time to seed a new data direction.
320 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
321 unsigned long long max_bytes;
325 * Put a top limit on how many bytes we do for
326 * one data direction, to avoid overflowing the
329 ddir = get_rand_ddir(td);
330 max_bytes = td->this_io_bytes[ddir];
332 (td->o.size * td->o.rwmix[ddir] / 100)) {
333 if (!td->rw_end_set[ddir]) {
334 td->rw_end_set[ddir] = 1;
335 fio_gettime(&td->rw_end[ddir], NULL);
341 if (ddir != td->rwmix_ddir)
344 td->rwmix_ddir = ddir;
346 return td->rwmix_ddir;
347 } else if (td_read(td))
353 static void put_file_log(struct thread_data *td, struct fio_file *f)
355 int ret = put_file(td, f);
358 td_verror(td, ret, "file close");
361 void put_io_u(struct thread_data *td, struct io_u *io_u)
363 assert((io_u->flags & IO_U_F_FREE) == 0);
364 io_u->flags |= IO_U_F_FREE;
367 put_file_log(td, io_u->file);
370 flist_del(&io_u->list);
371 flist_add(&io_u->list, &td->io_u_freelist);
375 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
377 struct io_u *__io_u = *io_u;
379 dprint(FD_IO, "requeue %p\n", __io_u);
381 __io_u->flags |= IO_U_F_FREE;
382 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
383 td->io_issues[__io_u->ddir]--;
385 __io_u->flags &= ~IO_U_F_FLIGHT;
387 flist_del(&__io_u->list);
388 flist_add_tail(&__io_u->list, &td->io_u_requeues);
393 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
395 if (td->io_ops->flags & FIO_NOIO)
399 * see if it's time to sync
401 if (td->o.fsync_blocks &&
402 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
403 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
404 io_u->ddir = DDIR_SYNC;
408 io_u->ddir = get_rw_ddir(td);
411 * See if it's time to switch to a new zone
413 if (td->zone_bytes >= td->o.zone_size) {
415 io_u->file->last_pos += td->o.zone_skip;
416 td->io_skip_bytes += td->o.zone_skip;
420 * No log, let the seq/rand engine retrieve the next buflen and
423 if (get_next_offset(td, io_u)) {
424 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
428 io_u->buflen = get_next_buflen(td, io_u);
430 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
434 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
435 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
436 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
437 io_u->buflen, io_u->file->real_file_size);
442 * mark entry before potentially trimming io_u
444 if (td_random(td) && file_randommap(td, io_u->file))
445 mark_random_map(td, io_u);
448 * If using a write iolog, store this entry.
451 dprint_io_u(io_u, "fill_io_u");
452 td->zone_bytes += io_u->buflen;
457 static void __io_u_mark_map(unsigned int *map, unsigned int nr)
486 void io_u_mark_submit(struct thread_data *td, unsigned int nr)
488 __io_u_mark_map(td->ts.io_u_submit, nr);
489 td->ts.total_submit++;
492 void io_u_mark_complete(struct thread_data *td, unsigned int nr)
494 __io_u_mark_map(td->ts.io_u_complete, nr);
495 td->ts.total_complete++;
498 void io_u_mark_depth(struct thread_data *td, unsigned int nr)
502 switch (td->cur_depth) {
524 td->ts.io_u_map[index] += nr;
527 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
564 assert(index < FIO_IO_U_LAT_U_NR);
565 td->ts.io_u_lat_u[index]++;
568 static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
609 assert(index < FIO_IO_U_LAT_M_NR);
610 td->ts.io_u_lat_m[index]++;
613 static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
616 io_u_mark_lat_usec(td, usec);
618 io_u_mark_lat_msec(td, usec / 1000);
622 * Get next file to service by choosing one at random
624 static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
631 long r = os_random_long(&td->next_file_state);
634 fno = (unsigned int) ((double) td->o.nr_files
635 * (r / (OS_RAND_MAX + 1.0)));
637 if (f->flags & FIO_FILE_DONE)
640 if (!(f->flags & FIO_FILE_OPEN)) {
643 err = td_io_open_file(td, f);
649 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
650 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
654 td_io_close_file(td, f);
659 * Get next file to service by doing round robin between all available ones
661 static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
664 unsigned int old_next_file = td->next_file;
670 f = td->files[td->next_file];
673 if (td->next_file >= td->o.nr_files)
676 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
677 if (f->flags & FIO_FILE_DONE) {
682 if (!(f->flags & FIO_FILE_OPEN)) {
685 err = td_io_open_file(td, f);
687 dprint(FD_FILE, "error %d on open of %s\n",
695 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
696 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
700 td_io_close_file(td, f);
703 } while (td->next_file != old_next_file);
705 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
709 static struct fio_file *get_next_file(struct thread_data *td)
713 assert(td->o.nr_files <= td->files_index);
715 if (td->nr_done_files >= td->o.nr_files) {
716 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
717 " nr_files=%d\n", td->nr_open_files,
723 f = td->file_service_file;
724 if (f && (f->flags & FIO_FILE_OPEN) && !(f->flags & FIO_FILE_CLOSING)) {
725 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
727 if (td->file_service_left--)
731 if (td->o.file_service_type == FIO_FSERVICE_RR ||
732 td->o.file_service_type == FIO_FSERVICE_SEQ)
733 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
735 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
737 td->file_service_file = f;
738 td->file_service_left = td->file_service_nr - 1;
740 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
744 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
749 f = get_next_file(td);
756 if (!fill_io_u(td, io_u))
760 td_io_close_file(td, f);
762 f->flags |= FIO_FILE_DONE;
764 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
771 struct io_u *__get_io_u(struct thread_data *td)
773 struct io_u *io_u = NULL;
775 if (!flist_empty(&td->io_u_requeues))
776 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
777 else if (!queue_full(td)) {
778 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
787 assert(io_u->flags & IO_U_F_FREE);
788 io_u->flags &= ~IO_U_F_FREE;
791 flist_del(&io_u->list);
792 flist_add(&io_u->list, &td->io_u_busylist);
800 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
801 * etc. The returned io_u is fully ready to be prepped and submitted.
803 struct io_u *get_io_u(struct thread_data *td)
808 io_u = __get_io_u(td);
810 dprint(FD_IO, "__get_io_u failed\n");
815 * from a requeue, io_u already setup
821 * If using an iolog, grab next piece if any available.
823 if (td->o.read_iolog_file) {
824 if (read_iolog_get(td, io_u))
826 } else if (set_io_u_file(td, io_u)) {
827 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
832 assert(f->flags & FIO_FILE_OPEN);
834 if (io_u->ddir != DDIR_SYNC) {
835 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
836 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
840 f->last_pos = io_u->offset + io_u->buflen;
842 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
843 populate_verify_io_u(td, io_u);
844 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
845 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
849 * Set io data pointers.
851 io_u->xfer_buf = io_u->buf;
852 io_u->xfer_buflen = io_u->buflen;
855 if (!td_io_prep(td, io_u)) {
856 if (!td->o.disable_slat)
857 fio_gettime(&io_u->start_time, NULL);
861 dprint(FD_IO, "get_io_u failed\n");
866 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
868 const char *msg[] = { "read", "write", "sync" };
870 log_err("fio: io_u error");
873 log_err(" on file %s", io_u->file->file_name);
875 log_err(": %s\n", strerror(io_u->error));
877 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
878 io_u->offset, io_u->xfer_buflen);
881 td_verror(td, io_u->error, "io_u error");
884 static void io_completed(struct thread_data *td, struct io_u *io_u,
885 struct io_completion_data *icd)
888 * Older gcc's are too dumb to realize that usec is always used
889 * initialized, silence that warning.
891 unsigned long uninitialized_var(usec);
893 dprint_io_u(io_u, "io complete");
895 assert(io_u->flags & IO_U_F_FLIGHT);
896 io_u->flags &= ~IO_U_F_FLIGHT;
898 if (io_u->ddir == DDIR_SYNC) {
899 td->last_was_sync = 1;
903 td->last_was_sync = 0;
906 unsigned int bytes = io_u->buflen - io_u->resid;
907 const enum fio_ddir idx = io_u->ddir;
910 td->io_blocks[idx]++;
911 td->io_bytes[idx] += bytes;
912 td->this_io_bytes[idx] += bytes;
914 if (ramp_time_over(td)) {
915 if (!td->o.disable_clat || !td->o.disable_bw)
916 usec = utime_since(&io_u->issue_time,
919 if (!td->o.disable_clat) {
920 add_clat_sample(td, idx, usec, bytes);
921 io_u_mark_latency(td, usec);
923 if (!td->o.disable_bw)
924 add_bw_sample(td, idx, bytes, &icd->time);
927 if (td_write(td) && idx == DDIR_WRITE &&
929 td->o.verify != VERIFY_NONE)
930 log_io_piece(td, io_u);
932 icd->bytes_done[idx] += bytes;
935 ret = io_u->end_io(td, io_u);
936 if (ret && !icd->error)
940 icd->error = io_u->error;
941 io_u_log_error(td, io_u);
945 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
948 if (!td->o.disable_clat || !td->o.disable_bw)
949 fio_gettime(&icd->time, NULL);
954 icd->bytes_done[0] = icd->bytes_done[1] = 0;
957 static void ios_completed(struct thread_data *td,
958 struct io_completion_data *icd)
963 for (i = 0; i < icd->nr; i++) {
964 io_u = td->io_ops->event(td, i);
966 io_completed(td, io_u, icd);
972 * Complete a single io_u for the sync engines.
974 long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
976 struct io_completion_data icd;
978 init_icd(td, &icd, 1);
979 io_completed(td, io_u, &icd);
983 return icd.bytes_done[0] + icd.bytes_done[1];
985 td_verror(td, icd.error, "io_u_sync_complete");
990 * Called to complete min_events number of io for the async engines.
992 long io_u_queued_complete(struct thread_data *td, int min_evts)
994 struct io_completion_data icd;
995 struct timespec *tvp = NULL;
997 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
999 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1004 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1006 td_verror(td, -ret, "td_io_getevents");
1011 init_icd(td, &icd, ret);
1012 ios_completed(td, &icd);
1014 return icd.bytes_done[0] + icd.bytes_done[1];
1016 td_verror(td, icd.error, "io_u_queued_complete");
1021 * Call when io_u is really queued, to update the submission latency.
1023 void io_u_queued(struct thread_data *td, struct io_u *io_u)
1025 if (!td->o.disable_slat) {
1026 unsigned long slat_time;
1028 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1029 add_slat_sample(td, io_u->ddir, io_u->xfer_buflen, slat_time);
1034 * "randomly" fill the buffer contents
1036 void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1037 unsigned int max_bs)
1039 long *ptr = io_u->buf;
1041 if (!td->o.zero_buffers) {
1042 while ((void *) ptr - io_u->buf < max_bs) {
1043 *ptr = rand() * GOLDEN_RATIO_PRIME;
1047 memset(ptr, 0, max_bs);